Cognitive Offloading and Competence Assessment in AI-Mediated Education

When AI contributes to learning products, output quality alone does not determine what can be claimed about the learner. This conceptual paper distinguishes product acceptance, opportunities to learn, and recognition of competence, together with individual, tool-supported, and offloading-governance targets. It critically synthesizes cognitive offloading, assessment validity, evaluative judgment, and AI assessment design. The main distinctions have substantial antecedents; the contribution is a six-step decision record that makes the target, conditions, learning pathway, diagnostic basis, learner evidence, and limits of recognition inspectable. Five conditional propositions constrain its use. A scheduling problem illustrates finite verification, and an open-ended policy brief examines incomplete checking and disagreement about goals, warrants, and values. Both cases are constructed analytical examples, not observations of learners or AI-model behavior. Vietnamese education documents are retained in a separate contextual supplement, not as tests of the framework. The decision aid specifies reasons to revise, narrow, or defer a claim without presuming that the competence being developed is already present. It has not been shown to improve learning, produce inter-rater agreement, or outperform existing frameworks. Its value is the explicit justification of task-level decisions under stated conditions. Release 1.3: English and Vietnamese editions. The package includes the 11-page main article and four-page contextual supplement in each language, both complete LaTeX source packages, and the Vietnamese vector figure sheet. The language editions report the same conceptual study. English is the primary metadata language. The work is a conceptual preprint, not an independently peer-reviewed article. Its constructed cases do not establish effectiveness, inter-rater reliability, or superiority over existing frameworks. AI assistance and source-reading limits are disclosed in the manuscripts. Authored manuscripts, supplements and original figures are licensed CC BY 4.0. The bundled IEEEtran class retains its original license notices. Raw third-party reference papers and the internal workspace snapshot are not distributed.